Show HN: Hexis, open-source Claude Skills management

What happened

I'll pull the Hexis repo and HN thread first so the paragraphs stay grounded in what's actually there.I'll write the paragraphs from the repo README and the HN numbers you gave, then count words so it stays inside 600–900.Hexis appeared on Hacker News as a Show HN from Bevel Software, titled open-source Claude Skills management. The repository is https://github.com/Bevel-Software/Hexis and the thread is https://news.ycombinator.com/item?id=49300784. It has 5 points and 0 comments. The item sits in the HN Claude, Codex, and Fable cluster. The README is broader than that title: Hexis is a git-backed control plane for AI-agent skills, tools, context, permissions, and identity. It is self-hosted, MCP-native, licensed Apache-2.0, and called the open-source core of the Bevel platform. A public demo runs at demo.bevel.software. Bevel also offers managed hosting, upgrades, backups, and SSO through ali.raza@bevel.software.

The catalogue is a git repository the operator owns. Skills, tool manuals, knowledge pages, and roles.yaml are files. Docker Compose starts Postgres and the app. First-run setup asks for the https clone URL of an empty knowledge-base repo, a token with contents read and write, and which branches are protected. Changes on protected branches land only after owners approve a change request. An agent that hits a broken skill mid-task can propose the fix; a person still decides. Agents connect over MCP as a remote Streamable HTTP server with OAuth. Named clients are Claude, Claude Code, Codex, Cursor, Cline, and ChatGPT. Discovery is on demand: list_skills and search narrow the field, get_skill returns one skill at call time so the context window is not flooded. Access is resolved per file against the signed-in user's role. The agent can only do what that person can do, and it never holds the credentials it uses. The tree is a pnpm monorepo with packages shared, core-backend, and core-frontend, plus apps server and web. Local development is Node 22, pnpm 10, git 2.41 or newer, and Postgres 17. A public box needs JWT_SECRET, SECRETS_ENC_KEY, PUBLIC_BACKEND_URL, and PUBLIC_FRONTEND_URL, and can sit behind a reverse proxy on port 3001.

The technical detail

The engineering problem is the instruction layer, not another model. Claude Skills are folders with a SKILL.md that teach an agent a class of tasks. Those folders, MCP tool configs, and company context currently scatter across laptops and vendor accounts. A team that runs Claude Code next to Codex or Cursor has no shared review path for the text those agents obey. Hexis treats that text like application code: commit history, blame, revert, and owner approval. Non-technical teammates can browse plugins and propose edits without touching client JSON. Per-file ACL lets a sales skill and an infra runbook share one workspace without sharing readers. The credentials claim matters to anyone who has watched an agent print a secret into a transcript. If you already keep prompts in a wiki, this is that wiki with a query API agents can call.

Why it matters for builders

MCP gateways already decide which tools an agent may invoke. Hexis states the difference itself: a gateway has no opinion on whether the skill telling the agent what to do is any good. Public SKILL.md lists collect reusable packages for Claude, Codex, Cursor, and other CLIs, but they are catalogues, not a review-gated control plane with SSO. Anthropic's Skills format is the unit Hexis manages. Bevel's paid offer sits on the Apache-2.0 core. The distribution bet is one MCP workspace and many clients, each scoped to the user's role. The Show HN title says Claude Skills. The README lists Codex, Cursor, Cline, and ChatGPT on the same connection path.

Market and competitive context

The path to try it is the one in the README. Clone the Hexis repo, copy .env.example, set ADMIN_EMAIL, ADMIN_PASSWORD, JWT_SECRET, and SECRETS_ENC_KEY, then start Docker Compose. Point the setup screen at an empty git repo on GitHub, GitLab, Bitbucket, Azure DevOps, or a self-hosted host. Connect Claude or Cline over MCP. Cline can install the public demo as a remote HTTP MCP server at https://demo.bevel.software/api/mcp and finish OAuth in the browser. The demo is shared and read-mostly: visitors propose, owners approve, and the fictional company knowledge is there to walk the loop. Watch whether the change-request queue stays usable when agents propose skill edits at volume. Watch whether Codex and ChatGPT stay first-class relative to a Claude-shaped title. Watch the thread itself. Five points and zero comments means almost nobody has argued this against a real internal knowledge base in public yet.

What to watch next

Self-hosting puts a git credential, a JWT secret, and a secrets encryption key on a machine every company agent will call. The never-holds-credentials claim depends on the MCP session and the role map, not on the model behaving. A five-minute Docker path plus an empty-repo seed will tempt teams to load real runbooks before SSO and backups are set. Open questions include how merge-heavy the propose-and-approve loop gets when agents file skill patches mid-task, whether per-file ACL plus git history is enough for shops that want a separate compliance export, and whether Apache-2.0 plus managed hosting is a split enterprises will accept if they will not run Compose themselves. Related prior art is the SKILL.md convention, remote MCP with OAuth, and the internal prompt libraries that never left a wiki. Hexis is that library with a review workflow and an MCP socket. It is not a new model and it does not replace a tool gateway.